Abstract
A quantitative understanding of the thermodynamics of biochemical reactions is essential for accurately modeling metabolism. The group contribution method (GCM) is one of the most widely used approaches to estimating standard Gibbs energies and redox potentials of reactions for which no experimental measurements exist. Previous work has shown that quantum chemical predictions of biochemical thermodynamics are a promising approach to overcome the limitations of GCM. However, the quantum chemistry approach is significantly more expensive. Here we use a combination of quantum chemistry and machine learning to obtain a fast and accurate method for predicting the thermodynamics of biochemical redox reactions. We focus on predicting the redox potentials of carbonyl functional group reductions to alcohols and amines, two of the most ubiquitous carbon redox transformations in biology. Our method relies on semi-empirical quantum chemistry calculations calibrated with Gaussian Process (GP) regression against available experimental data. Our approach results in higher predictive power than the GCM at a low computational cost. We design and implement a network expansion algorithm that iteratively reduces and oxidizes a set of natural seed metabolites, and demonstrate the high-throughput applicability of our method by predicting the standard potentials of more than 315,000 redox reactions involving approximately 70,000 compounds. Additionally, we developed a novel fingerprint-based framework for detecting molecular environment motifs that are enriched or depleted across different regions of the redox potential landscape. We provide open access to all source code and data generated.
Introduction
To model metabolism or design new synthetic biological pathways, scientists need to know the "redox potential"—the electrical pressure required to move electrons—of biochemical reactions. However, experimental data exists for only about 10% of known natural reactions. Traditionally, researchers use the Group Contribution Method (GCM), which estimates energy by summing values of a molecule’s functional groups. While fast, GCM is "blind" to complex intramolecular interactions, leading to inaccuracies. Conversely, high-level Quantum Chemistry (QC) can be very accurate but is computationally expensive, taking hours or days for a single reaction. This paper presents a hybrid framework that combines the speed of semi-empirical quantum chemistry with the corrective power of machine learning to map the thermodynamic landscape of biochemistry at an unprecedented scale.
Computational Pipeline
The authors developed a multi-step pipeline to bridge the gap between speed and accuracy. They used a fast, semi-empirical quantum method called PM7 to calculate the electronic energies of molecules. Raw electronic energies were then converted into biochemical potentials ($E'^{m}$ at pH 7) using chemoinformatic pKa estimates and the Alberty-Legendre transform.
The core innovation is the use of Gaussian Process (GP) regression to "fix" systematic errors in the cheap PM7 calculations. To inform the GP model, the researchers used "reaction fingerprints"—mathematical vectors representing the structural difference between the product and the substrate. This allows the model to learn the energetics of the transformation itself.
Network Expansion and Large-Scale Application
To demonstrate the high-throughput applicability of their method, the authors implemented a network expansion algorithm. Starting with "seed" metabolites from the KEGG database, the algorithm iteratively oxidizes and reduces molecules to generate a massive network of non-natural reactions.
This process generated a network of 315,000 redox reactions involving approximately 70,000 compounds, all derived from natural metabolites with 21 or fewer carbon atoms. This represents the largest existing database of biochemical redox potentials.
Key Results and Insights
- Accuracy: The PM7+GP approach achieved a Mean Absolute Error (MAE) of 23.5 mV, significantly better than the standard GCM (27.3 mV) and raw PM7 (30.4 mV).
- Efficiency: The calibrated PM7 method takes only ~0.37 seconds of compute time per reaction, compared to over 17 hours for high-level coupled-cluster methods.
- Global Mapping: The study found that carbonyl-to-alcohol reductions typically peak at ~ -220 mV, while carbonyl-to-amine reductions peak at ~ -235 mV.
- Structural Motifs: By binning reactions by predicted potentials and averaging fingerprints, the authors identified structural motifs (like dicarbonyls in heterocyclic rings) that correlate with specific energy levels.
Conclusion
This study proves that a "cheap" theory (PM7) plus a smart ML correction (GP) can match the accuracy of gold-standard quantum methods while being orders of magnitude faster. While currently focused on carbonyl reductions due to data availability, this framework could be used to fine-tune the energetics of synthetic metabolites for metabolic engineering or to predict drug metabolism. The code and the 315k reaction dataset are available at: https://github.com/aspuru-guzik-group/gp_redox_rxn
Summary updated